Document-over-topic probabilities don't sum to unit for sparsity regularization.

[KVasya]

When using artm.SmoothSparseThetaRegularizer(tau=tau_val) with tau_val<0 we get some \Theta matrix columns filled totally with zeros. From perplexity score, the optimization converges. The quantity of documents with all zeros in their \Theta columns grows as $tau_val-&gt;-\infty$.
How it's possible that optimization constraint on theta columns violates?